Field log — Oct 2026Temporal code memoryIn build
Copilot tells you what the code does. We're building the why.
StrataCode imports your repo into a timestamped graph — every fix, revert, and PR discussion linked to the lines it touched. Ask “why is this like this?” about any date, get a cited timeline back. When the evidence is thin, it tells you so and lists exactly what to fetch next — never a made-up answer.
Per-tenant quotas + BYOK · questions? hello@novare.systems
weeks → minutes
onboarding goal
Design target
history_ref
on every finding
In build
≥0.9 / ≥0.95
citation · era gates
Design target
Proven pieces, no reinvention
ask-head · why-QA
$ stratacode ask "Why does auth retry 3 times?" --as-of 2024-06-14
2023-11-02 · PR #418 · retry 1 → 3
Flaky IdP timeouts during deploys. Bumped to 3 with backoff.
2024-02-19 · issue #502 · alert storm
Kept at 3, added jitter after on-call flagged the herd.
2024-06-14 · as_of ✓ · commit 9f3a…
Held then. Re-check after the IdP migration.
Answer cites 8/8 claims · NLI-verified
PR #418issue #502commit 9f3aDesign rule: >40% uncited ⇒ abstains + fetch-plan, no hallucination.
01The problem
Copilot explains the code. Nobody explains the history.
Built for solo developers and small agencies living with onboarding and audit pain — where losing one person's memory means losing the project's memory.
Onboarding takes weeks
Every new developer re-asks the same history questions — and the answers live in someone's head, or in a PR thread from two years ago nobody can find.
Refactors revert old fixes
Without the rationale attached, a well-meaning cleanup removes the weird-looking line that was load-bearing. The bug comes back. Nobody connects the two.
Reviews miss the context
Approvers see the diff, not the debate. The old decision that motivated the current shape is invisible at review time — so it gets re-litigated or ignored.
02Product In build
One graph, two heads, one flywheel.
Ask-head answers why with citations. Review-head reviews PRs with the past attached. The learner keeps what merges teach. All three read the same temporal graph — below is the designed flow, shown with illustrative fixtures.
Ask “why is this like this?” about any date
In buildIllustrativeq + as_of → embed → Qdrant top-30 (ts ≤ as_of) → 2-hop graph → rerank top-8 → LangGraph chain → NLI verifier → answer + timeline + verdict + fetch-plan
PR #418issue #502commit 9f3aera: pre-migration03How it works
GitHub events in. Timestamped memory out.
Webhooks are verified at the edge, imports run as resumable jobs, and everything lands in one graph with two stores: TiDB for timestamped nodes and edges, Qdrant for time-filtered vectors — plus era summaries a human can skim.
GitHub events → edge HMAC verify → API gateway → Trigger.dev jobs
→ TEMPORAL GRAPH (TiDB nodes/edges + Qdrant vectors + era summaries)
→ ask-head · review-head · learner (all read the same graph)
Verified ingestion
Signatures checked at the edge, events deduped by hash, imports checkpointed so big repos resume instead of restarting.
TiDB: the who and when
Commits, PRs, issues, people, and edges (introduced_by, fixed_by, discussed_in, reverts, touches) — every row timestamped.
Qdrant: the what, time-filtered
Chunk vectors filtered to ts ≤ as_of, so asking about June never leaks August's answer. Free-tier budget: ~50k chunks per repo, 5 pilot repos.
Strict BYOK + quotas
Members use their own provider keys or calls fail closed. Per-tenant daily token quotas, budget alerts, and a kill-switch that halts AI spend fast.
Abstention is a feature
The NLI verifier drops uncited claims. Over 40% dropped means no answer — just unknown plus a fetch-plan. Honest beats fluent.
$0 to start, no servers
Vercel + TiDB Cloud + Qdrant Cloud + Upstash + Trigger.dev free tiers. No Docker, no self-hosting, clone-to-healthy in minutes.
01 — Cited or it didn't happen
Every sentence in an answer points at a commit, PR, or issue. No pointer, no sentence.
02 — Abstain over hallucinate
Unknown with a fetch-plan beats a confident fabrication. The verifier has the final say, not the generator.
03 — Memory compounds
Each merge teaches era patterns, so the tenth question about retries starts from nine answers, not zero.
Why not just use…?
| Tool | Answers | Remembers why |
|---|---|---|
| Copilot-style assistants | What the code does | No — no history graph |
| Code search + blame | Where lines came from | Partly — lines, not rationale |
| PR review bots | What looks risky now | No — no memory of past decisions |
| StrataCode In build | What + whether it still holds | Yes — timestamped rationale graph + abstention + eval gates |
Eval gates every deploy
Design targetFrozen sets — 100 why-questions, 50 review cases — must pass before anything ships. Drift blocks deploy, cost per ask is tracked against a $0.02 target, and retrieval floors are already enforced in config with a scripted-harness baseline around precision ~0.72 / recall ~0.98 — published honestly, because a gate you can see is a gate you can trust.
0
+ 50 review cases
frozen sets
≥0.0
citation precision
deploy gate
≥0.00
era accuracy
deploy gate
<$0.00
cost per ask
tracked target
Nightly eval cron · abstention calibration reported · user data deletable on repo disconnect
04Pricing
Start free. Pay when memory pays.
The pilot is free while we build in the open — 5 repos fit the free-tier capacity exactly. Paid tiers are planned around per-tenant quotas and your own keys, so our incentives stay simple: useful memory, metered fairly.
Pilot
LiveStart here — openings now
Free
For the first believers
- Up to 5 repos (free-tier capacity)
- Ask-head why-QA with timelines
- Per-tenant quotas included
- Direct line to the builder
Team
Design targetPlanned
Per-tenant quotas + BYOK
- Everything in Pilot, more repos
- Review-head PR memory
- Bring your own provider keys
- Usage + cost dashboard
Enterprise
Design targetPlanned
Audit-grade memory
- Everything in Team
- Audit log + retention controls
- SSO and data processing terms
- Priority support
05Model roadmap Design target
Today: small open models. Next: Claude as the reasoning engine.
Honest status, because this page exists for the Claude for Startups reviewers too: StrataCode reasons today with Llama 3.1 8B and Qwen 2.5 Coder behind a provider interface (NVIDIA NIM → Groq → OpenRouter fallback). There is no Claude or Anthropic code in the product yet — the interface is model-agnostic, the eval harness is frozen, and the application goal is to make Claude the primary reasoning engine for the ask-chain, the NLI verifier, and the review agents, judged by the same ≥0.9 / ≥0.95 gates.
Now
Llama 3.1 8B / Qwen 2.5 Coder via NIM → Groq → OpenRouter. Jina v4 embeddings + reranker. Free-tier infra, strict BYOK.
With Claude
Claude plans before code, verifies before answering, reviews with judgment — same abstention rule, harder model behind it.
You can check
Provider order, quotas, and eval floors live in versioned config. Ask us anything at hello@novare.systems — a human replies.
—A note from the builder
I'm building StrataCode because I've watched the same history get lost twice — once when the author left, and once when a tidy refactor reverted their fix. The code survived. The reason didn't.
So this site makes exactly one promise it can keep: everything here is labeled — what's live, what's in build, what's a design target. No fake customers, no borrowed logos. If you have a repo whose history is worth keeping, write to me directly and we'll import it together.
Vivek ChudasamaNovare Systems · v0.2.0-drafthello@novare.systems
06FAQ
Asked like a human, answered like one.
No — and we won't pretend otherwise. Reasoning runs on Llama 3.1 8B and Qwen 2.5 Coder through a provider interface with fallbacks. Claude is our stated next step: we're applying to Claude for Startups to evaluate Claude as the primary reasoning engine behind the same frozen eval gates.
Still stuck? hello@novare.systems
Your repo already knows why. Let's make it tell you.
Join the pilot — first 5 repos free, no servers, no sales call. Just history with receipts.
01 · Week 1
Connect repos, import history, ask your first why.
02 · Week 2
Review PRs with the past attached to every finding.
03 · Ongoing
Merges teach era patterns — memory compounds.
hello@novare.systems · replies from a human